Qwen Image API: Integration Guide, Pricing & Code

The Qwen image API is one model family with a confusing name ladder. qwen-image, qwen-image-plus, qwen-image-max, qwen-image-2.0, qwen-image-2.0-pro, and now qwen-image-3.0 and 3.0-pro all describe different tiers of the same text-to-image engine, and the official pricing for them sits on a separate page with five region tabs. If you just want to know which model to call, what it costs today, and how to run it, the official documentation makes you assemble that answer from three or four pages yourself.

This guide does it in one pass: the model lineup, official prices read live on August 23, 2026, working code for both the direct route (Alibaba Cloud Model Studio) and a unified-API route (Modellix), and the five operational pitfalls the docs bury. Modellix is an API aggregator and has a commercial interest in the second route, so prices on both sides are shown and the comparison stays like-for-like.

The Qwen image API model lineup in 2026

Qwen image generation exists in four generations, and the API exposes most of them side by side. The naming tells you the tier, not the age: plus and max are older, cheaper-quality tiers of the original model; 2.0 and 2.0-pro are the second generation; 3.0 and 3.0-pro are the current release (announced July 20, 2026); the edit variants are image-editing models. One more wrinkle: the official API reference pins each model to a dated snapshot ID (for example qwen-image-2.0-pro-2026-06-22), so the same model name can map to several versioned releases over time, and the prices in this guide track the snapshots current on the access date.

Model IDs and positioning from the official Qwen-Image API reference, accessed August 23, 2026.

What makes the whole family distinctive is in-image text rendering: posters, infographics, UI mockups, and multi-language layouts where the text inside the image has to be legible and correctly spelled. The 3.0 generation pushes this further: up to 4.5K token prompts, text as small as 10px, and native rendering in 12 languages, per the Qwen-Image-3.0 launch post. The open-weight lineage (Apache 2.0 for the original) is on GitHub, which is why so many platforms host it.

For a new integration, start with qwen-image-3.0 or qwen-image-3.0-pro. The 2.0 family remains relevant when you need batch output (n up to 6) at a fixed per-image price, and the edit models when your task is image-to-image transformation rather than generation.

Qwen image API pricing

Official pricing is per successfully generated image, billed on the Alibaba Cloud Model Studio page; failed calls cost nothing and do not consume free quota. The Singapore (International) region is the one to use without a Chinese mainland account; prices there differ slightly from the Beijing region.

Official Model Studio pricing for Qwen image models (Singapore tab), captured from the live page on August 23, 2026. The 3.0 series adds a $0.003/image input charge and free quotas of 10 images; the text-to-image models offer 100 free images.

Sources: Model Studio model pricing, accessed August 23, 2026, and the Modellix model pages, re-read September 6, 2026. For the provider's full catalog, see the Alibaba API pricing guide.

Two caveats before you compare numbers. First, the official 3.0-series price depends on output resolution (1K vs 2K) and charges for input images on editing calls; Modellix displays a single per-image rate. Second, these prices move, so treat this table as a dated snapshot, not a quote. Read it per row, because the comparison cuts both ways once you add the official input fee: for a single-input 3.0 Pro call at 1K, official is $0.043 against Modellix's $0.0675, while at 2K official Pro is $0.078 against $0.0675, and official 3.0 is $0.033 against Modellix's $0.0270. The cheaper route depends on the resolution you generate at. Modellix no longer lists the 2.0 series (its model pages returned "Model Not Found" on September 6, 2026), so 2.0 is a direct-route model now.

Calling the Qwen image API directly (Alibaba Cloud Model Studio)

You need three things: a Model Studio account, an API key, and the workspace ID embedded in the endpoint. The official Qwen-Image API reference covers key setup and the full parameter list; note that Beijing and Singapore keys and endpoints are separate and cannot be mixed.

The synchronous endpoint for the 3.0 series (Singapore):

curl --location 'https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation' \
--header 'Content-Type: application/json' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--data '{
    "model": "qwen-image-3.0-pro",
    "input": {
      "messages": [
        {
          "role": "user",
          "content": [
            { "text": "A vertical outdoor portrait with warm afternoon light. A newsstand on a sunlit street, dark green vines over the eaves, the sign reads \"Il Messaggero\" in white Gothic lettering." }
          ]
        }
      ]
    },
    "parameters": {
      "prompt_extend": true
    }
}'

The Python equivalent uses the DashScope SDK:

import os
import dashscope
from dashscope import MultiModalConversation

dashscope.base_http_api_url = "https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/api/v1"  # Singapore region; replace {WorkspaceId}

response = MultiModalConversation.call(
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    model="qwen-image-3.0-pro",
    messages=[{"role": "user", "content": [{"text": "A sunlit street portrait with a newsstand and a sign reading Il Messaggero"}]}],
    result_format="message",
    stream=False,
)

if response.status_code == 200:
    image_url = response.output.choices[0].message.content[0]["image"]
    print(image_url)  # download this now: it expires in 24 hours

Both examples follow the official API reference above. The response returns a temporary image URL (PNG) that expires after 24 hours, so download it to your own storage immediately. Only qwen-image and qwen-image-plus support the asynchronous task workflow (send X-DashScope-Async: enable, poll the returned task_id); the 2.0 and 3.0 series are called synchronously. The 3.0 series has its own API reference with additional regions (Hong Kong, Frankfurt, Tokyo) and the I2I/editing request shape.

Calling Qwen image through Modellix: one key, the whole family

Modellix exposes the current Qwen image generation (3.0, 3.0-pro, and the 3.0 edit models) behind one key and one async pattern; the 2.0 series and the legacy edit tiers are no longer listed. The endpoint follows the https://api.modellix.ai/api/v1/<provider>/<model> convention, so switching from qwen-image-3.0-pro to qwen-image-3.0-pro-edit is a one-line change:

curl --request POST \
  --url https://api.modellix.ai/api/v1/alibaba/qwen-image-3.0-pro \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "prompt": "A sunlit street portrait with a newsstand and a sign reading Il Messaggero",
    "size": "2048*2048",
    "prompt_extend": true
  }'

Media generation is asynchronous, so the first call returns a task handle, not an image:

{
  "code": 0,
  "message": "success",
  "data": {
    "status": "pending",
    "task_id": "task-qwen-img-123",
    "get_result": {
      "method": "GET",
      "url": "https://api.modellix.ai/api/v1/tasks/task-qwen-img-123"
    }
  }
}

Poll the get_result.url until status becomes success, then download the output. The pattern is documented in the Modellix API docs and per-model in the Qwen Image 3.0 Pro reference. If you prefer not to poll, send an X-Webhook-URL header and Modellix POSTs the result when the task finishes.

Every Qwen image API call on the Modellix route follows the same async pattern: submit a task, poll the task id, then download the result. The webhook option replaces step two.

Why would you pick the aggregator route over direct? The honest reasons: one key for Qwen and the rest of the platform's 210+ image and video models, no Chinese mainland account needed, per-call cost logs, and pay-as-you-go billing with no prepaid packages. The honest boundary: for a single model at high volume, direct Alibaba billing can be cheaper per image; the operational value of one key and one bill is what you are paying for, not a guaranteed lower unit price.

Qwen image API parameters that actually matter

The parameter set is small, and four knobs change your results more than the rest:

• prompt_extend: defaults to true and silently rewrites your prompt for more detail. Leave it on for diverse output; turn it off when you need tight control over a specific composition. The 3.0 series adds prompt_extend_mode (direct / agent).
• negative_prompt: up to 500 characters, only on models that support it (the 2.0/3.0 Pro tiers do). Useful for excluding "low resolution, distorted hands, warped text" style artifacts.
• size: the 2.0 series accepts any width*height between 512×512 and 2048×2048 total pixels (recommended presets include 2688*1536 for 16:9 and 2368*1728 for 4:3); max/plus accept a fixed set ending at 1664*928.
• n: number of images per call. 1 to 6 on the 2.0/3.0 series; fixed at 1 on max and plus; any other value returns an error.

Prompt length limits differ by model: 1,300 tokens on the 2.0 series, 800 on most others, and about 4,500 tokens on 3.0 (which is what makes dense layouts like newspapers and storyboards possible). Excess tokens are silently truncated, so if your complex prompt keeps losing detail, check the cap before blaming the model.

Common Qwen image API errors and how to fix them

1. Region mismatch (authentication or service errors). A Beijing key against a Singapore endpoint fails. Keys and endpoints are region-scoped, so pick one region and keep both in sync.
2. "current user api does not support synchronous calls". You called the async-only models (qwen-image, qwen-image-plus) without the X-DashScope-Async: enable header, or used an HTTP endpoint that requires it. Add the header and poll the task, or switch to a synchronous model.
3. "n must be 1". You requested n > 1 on qwen-image-max or qwen-image-plus, where batch output is not supported.
4. The image URL is dead by the time you need it. Result URLs and task data expire after 24 hours. Download to your own storage (S3, R2, your server) at generation time, not when you deploy.
5. Single-turn restriction. The messages array must contain exactly one user message; multi-turn conversations are not supported. Wrap each generation in its own call.

Qwen Image 2.0: Snapshot IDs, Status, and When to Stay on It

Qwen-Image-2.0 launched in February 2026 as a single model that merged generation and editing, with native 2K output and prompts up to about 1,300 tokens. Since Qwen Image 3.0 arrived in July 2026 it is the previous generation: still listed and priced on Model Studio without a legacy label, but no new 2.0 snapshot has shipped since June 2026. Treat it as frozen but served.

The bare IDs are aliases for dated snapshots:

Call the bare ID and you get whatever snapshot the alias tracks; call a dated ID when output must not change silently. Because 2.0 unifies generation and editing, an edit call costs the same $0.035 or $0.075 as a generation. The Beijing region ({WorkspaceId}.cn-beijing.maas.aliyuncs.com) has its own prices and no free quota. Two more 2.0 parameters: watermark defaults to false, and seed (0 to 2147483647) gives similar, not identical, output on repeat.

When does staying on 2.0 make sense? When your pipeline and prompts are already tuned to it (switching models is a regression-testing project), when you want a flat per-image price with no input-image fee or resolution tier, or when your work is batch posters and product visuals that never need 3.0's long prompts and small multilingual text. For new integrations, infographics, and localized layouts, start on 3.0. Modellix no longer lists the 2.0 series, so 2.0 means the direct Model Studio route.

Qwen image API free quota and free tiers

Model Studio offers a free quota only in the Singapore region: 100 images for most text-to-image models and 10 for the 3.0 series, valid for 90 days from account activation. Other regions have no free quota. Modellix has no automatic signup credit; Qwen image calls are pay-as-you-go. For a wider model comparison, see our text-to-image API guide.

FAQ

What is the Qwen image API?

The Qwen image API is Alibaba's hosted interface to the Qwen-Image family of text-to-image and image-editing models, available through Alibaba Cloud Model Studio (DashScope) and through aggregators such as Modellix. Its signature capability is accurate text rendering inside generated images: posters, infographics, and UI layouts where the text has to be legible and correctly spelled.

Which Qwen image model should I use?

For new text-to-image work, use qwen-image-3.0-pro when photorealism and complex layouts matter, qwen-image-3.0 when you want the best quality/speed balance, and qwen-image-2.0 or 2.0-pro when you need batch output (1 to 6 images per call) or strict negative-prompt control at a flat per-image price. Use the edit variants for image-to-image transformation.

How much does the Qwen image API cost?

Official Singapore-region prices are $0.03 to $0.075 per successfully generated image depending on model and resolution, with the 3.0 series adding a $0.003/image input charge. Modellix displays a single per-image rate per model (for example $0.0675/img for qwen-image-3.0-pro). Failed calls are not billed on either side.

Is there a free Qwen image API?

Model Studio grants a free quota in the Singapore region only: 100 images for most models or 10 for the 3.0 series, valid 90 days. Modellix has no automatic signup credit; its Qwen image calls are billed per image.

Can I use the Qwen image API without a Chinese account?

Yes. The Singapore (International) region of Model Studio is open to international accounts and uses its own API key. Do not mix Singapore and Beijing keys; they are region-scoped and cross-region calls fail.

What is the difference between qwen-image and qwen-image-3.0?

qwen-image is the original 2025 generation (Apache 2.0, open weights); qwen-image-3.0 is the current generation with much longer prompt support (4.5K tokens), finer text rendering (down to 10px), 12-language rendering, and stronger world knowledge for UI and infographic output. The 3.0 series also merges generation and editing into one model.

Can I edit images with the Qwen image API?

Yes. Dedicated editing models (qwen-image-edit, edit-plus, edit-max, qwen-image-2.0-edit, qwen-image-2.0-pro-edit, and the 3.0 edit variants) perform instruction-based edits and image fusion; the 3.0 generation and editing models also accept 1 to 3 input images with an edit instruction.

Is the Qwen image API OpenAI-compatible?

The Qwen image models are not OpenAI-compatible; they use the DashScope native multimodal-generation endpoint. OpenAI-compatible mode applies to Qwen text models (qwen-plus and up), which is how agent frameworks orchestrate image generation as a tool call.